With transforEmotion you can use cutting-edge transformer models for zero-shot emotion classification of text, image, and video in R, all without the need for a GPU, subscriptions, paid services, or using Python.
if(!"devtools" %in% row.names(installed.packages())){
install.packages("devtools")
}
devtools::install_github("atomashevic/transforEmotion")After installing the package, load it in R.
After loading package for the first time, you need to setup the Python virtual environment. This will download the necessary Python packages and models. This step can take a few minutes but it is only required once after installing the package on a new system.
Next load some data with text for analysis. The example below uses item descriptions from the personality trait extraversion in the NEO-PI-R inventory found on the IPIP website.
For the example, the positively worded item descriptions will be used.
Next, the text can be loaded in the function transformer_scores() to obtain the probability that item descriptions correspond to a certain class. The classes defined below are the facets of extraversion in the NEO-PI-R. The example text data draws from the friendliness facet.
# Cross-Encoder DistilRoBERTa
transformer_scores(
text = text,
classes = c(
"friendly", "gregarious", "assertive",
"active", "excitement", "cheerful"
)
)The default transformer model is DistilRoBERTa. The model is fast and accurate.
Another model that can be used is BART, a much larger and more computationally intensive model (slower prediction times). The BART model tends to be more accurate but the accuracy gains above DistilRoBERTa are negotiatiable.
# Facebook BART Large
transformer_scores(
text = text,
classes = c(
"friendly", "gregarious", "assertive",
"active", "excitement", "cheerful"
),
transformer = "facebook-bart"
)Text classification models with a pipeline on huggingface can be used so long as there is a pipeline available for them. Below is an example of Typeform’s DistilBERT model.
# Directly from huggingface: typeform/distilbert-base-uncased-mnli
transformer_scores(
text = text,
classes = c(
"friendly", "gregarious", "assertive",
"active", "excitement", "cheerful"
),
transformer = "typeform/distilbert-base-uncased-mnli"
)For Facial Expression Recognition (FER) task from images we use Open AI’s CLIP transformer model. Two input arguments are needed: the path to image and list of emotion labels.
Path can be either local or an URL. Here’s an example of using a URL of Mona Lisa’s image from Wikipedia.
# Image URL or local filepath
image <- "https://upload.wikimedia.org/wikipedia/commons/thumb/7/76/Leonardo_da_Vinci_-_Mona_Lisa.jpg/401px-Leonardo_da_Vinci_-_Mona_Lisa.jpg"
# Array of emotion labels
emotions <- c("excitement", "happiness", "pride", "anger", "fear", "sadness", "neutral")
# Run FER
image_scores(image, emotions)You can define up to 10 emotions. The output is a data frame with 1 row and columns corresponding to emotions. The values are FER scores for each emotion.
If there is no face detected in the image, the output will be a 0x0 data frame.
If there are mulitple faces detected in the image, by default the function will return the FER scores for the larget (focal) face. Alternative is to select the face on the left or the right side of the image. This can be done by specifying the face_selection argument.
Video processing works by extracting frames from the video and then running the image processing function on each frame. Two input arguments are needed: the path to video and list of emotion labels.
Path can be either local filepath or a YouTube URL. Support for other video hosting platforms is not yet implemented.
# Video URL or local filepath
video_url <- "https://www.youtube.com/watch?v=hdYNcv-chgY&ab_channel=Conservatives"
# Array of emotion labels
emotions <- c("excitement", "happiness", "pride", "anger", "fear", "sadness", "neutral")
# Run FER on `nframes` of the video
result <- video_scores(video_url, classes = emotions,
nframes = 10, save_video = TRUE,
save_frames = TRUE, video_name = 'boris-johnson',
start = 10, end = 120)Working with videos is more computationally complex. This example extracts only 10 frames from the video and I shouldn’t take longer than few minutes on an average laptop without GPU (depending on your internet connection needed to download the entire video and CLIP model). In research applicatons, we will usually extract 100-300 frames from the video. This can take much longer, so pantience is advised while waiting for the results.
Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., … & Zettlemoyer, L. (2019). Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461.
Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., … & Stoyanov, V. (2019). Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.
Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., & Sutskever, I. (2021). Learning Transferable Visual Models From Natural Language Supervision. arXiv preprint arXiv:2103.00020
Yin, W., Hay, J., & Roth, D. (2019). Benchmarking zero-shot text classification: Datasets, evaluation and entailment approach. arXiv preprint arXiv:1909.00161.